On September 16, 2026, data-center specialists said calls to slow frontier-AI development had not produced an immediate change in demand. The reason is straightforward: training the most advanced models is only one part of the market. Cloud computing, enterprise workloads, AI inference and projects already financed or under construction continue to keep the pipeline moving.
The comments followed a call by Anthropic CEO Dario Amodei on September 12 to slow the pace of frontier-model capability improvements. Sam Altman and Elon Musk endorsed the direction, while Demis Hassabis described it as appropriate for the moment. The discussion concerns the pace of frontier-model development—not a confirmed industry-wide construction pause.
Why a frontier-AI slowdown would not automatically stop data-center expansion
A data center is not built for just one task or one model. Its servers can support cloud services, business software, storage, conventional computing and multiple AI workloads. That broader customer base makes a slowdown in frontier-model training less likely to translate immediately into a sector-wide contraction.
John Andril, a critical-infrastructure specialist at Avison Young, said on September 16 that he saw no current effect on demand. Anthony Wanger of Regnaw Capital likewise characterized the proposals as a way to pace model development rather than stop it altogether.
A coordinated global pause would also be difficult to organize. The specialists said it would require competing companies to comply at the same time, including companies in China. A voluntary restraint by some laboratories would not automatically bind their rivals.
The part of demand most exposed to a training slowdown
Frontier-model training is the workload most directly exposed to a decision to slow capability development. Andril estimated that it represents 10% to 20% of current and projected data-center demand. That is an estimate attributed to Andril, not a measured industry-wide share.
That estimate also describes the potential exposure of the broader market, not a forecast that the same percentage of facilities would become unused. Data-center demand includes several segments with different customers, contracts and operating schedules.
| Demand segment | Relationship to a frontier-training slowdown | Near-term implication |
| Frontier-model training | Most directly exposed to a slower pace of capability development | Facilities designed specifically for this workload would face the clearest risk |
| AI inference | Runs trained models for users and requires capacity closer to end users | Continued AI adoption can keep adding demand for distributed capacity |
| Cloud computing | A major demand driver separate from frontier-model training | Cloud expansion can continue even if frontier training slows |
| Enterprise workloads | A continuing part of the data-center market | Business computing supports demand beyond AI laboratories |
| Financed or under-construction capacity | Projects are committed months or years before facilities become operational | Existing work is less sensitive to an immediate change in model-training plans |
Cloud, enterprise workloads and inference keep the market broader than model training
Inference is the stage at which a trained model produces answers or performs tasks for users. Unlike a large training run that may be concentrated in a limited number of facilities, inference can require computing capacity distributed closer to customers. That creates a separate infrastructure need as more people and businesses use AI services.
Andrew Power, CEO of Digital Realty, said on September 15 that cloud-computing growth and digital transformation unrelated to AI remained major drivers of his company’s business. He described a potential AI slowdown as something other than a signal to stop building or investing across the data-center sector.
The distinction matters for anyone reading “AI slowdown” as shorthand for “less demand for computing.” A slower release schedule for frontier models and a broad decline in cloud, enterprise or inference workloads are different events. The first could occur without automatically causing the second.
Why the effect could take time to appear
Most new data-center capacity is committed months or years before a facility becomes operational. Specialists said much of the infrastructure needed to support AI over the next five to 10 years is already financed or under construction.
That long project cycle creates a lag between a change in corporate plans and a visible change in construction activity. A decision affecting future training capacity would first have to alter contracts, financing, equipment orders or development schedules before it appeared in the physical pipeline.
Supply constraints and a large backlog add another layer. Andril said the backlog could persist even without demand from frontier-model training, and that a somewhat slower construction pace might ease pressure on developers rather than eliminate the need for new facilities.
Forecasts still point to expansion, with wide scenario ranges
The forward-looking numbers remain large, although they describe forecasts rather than completed results. A global forecast puts data-center demand at about 92 GW by 2027, with growth scenarios ranging from about 14% to 20%; its baseline scenario uses a 17% compound annual growth rate between 2025 and 2028.
A separate U.S. estimate projects potential AI-data-center power demand rising from 4 GW in 2024 to 123 GW by 2035. That outlook also identifies grid capacity, supply chains, permitting and demand forecasting as constraints. Some U.S. grid-connection requests can face waits of up to seven years.
The forecasts cover different geographies, time periods and methods, so they are not a single combined prediction. Together, they show why a few public calls for restraint have not translated into an immediate reversal of data-center expansion.
The near-term picture is therefore one of continued buildout alongside a debate about how quickly frontier models should improve. The most exposed slice is frontier training; cloud services, enterprise computing, inference and committed infrastructure keep the rest of the market moving.
